From 4bec99ecd07663f2049d860912531a9a85ccc4e1 Mon Sep 17 00:00:00 2001 From: Yusong Gao Date: Sat, 2 Aug 2025 14:43:45 +0800 Subject: [PATCH] Fix: resolve prefill of retracted request out-of-memory issue when ignore_eos is enabled (#7434) --- python/sglang/srt/managers/schedule_policy.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/python/sglang/srt/managers/schedule_policy.py b/python/sglang/srt/managers/schedule_policy.py index c07df2150..eb14b9835 100644 --- a/python/sglang/srt/managers/schedule_policy.py +++ b/python/sglang/srt/managers/schedule_policy.py @@ -455,7 +455,9 @@ class PrefillAdder: if not self.is_hybrid: # Skip this logic for swa. The SWA has different memory management, and # this mechanism is underestimating the memory usage. - cur_rem_tokens = self.cur_rem_tokens - len(req.origin_input_ids) + cur_rem_tokens = self.cur_rem_tokens - self.ceil_paged_tokens( + req.extend_input_len + ) tokens_freed = 0 for i, (tokens_left, tokens_occupied) in enumerate(self.req_states): # tokens_left gives a reservative calculation as the last token is not stored